Cross entropy gradient analysis for Alzheimer's disease assessment

Abstract

Background

Alzheimer's disease (AD) involves progressive cognitive decline associated with disrupted coordination and information exchange across brain regions. Cross-entropy can characterize inter-regional information flow, but its role in AD remains unexplored.

Objective

To evaluate cross-regional brain entropy (CRBEN) derived from resting-state fMRI in normal controls (NC), mild cognitive impairment (MCI), and AD, and to assess its potential as a biomarker of disease progression.

Methods

Resting-state fMRI data from 40 NC, 38 MCI, and 40 AD participants from ADNI 2/3 were preprocessed using SPM12 and FSL, including motion correction (FD < 0.5 mm). Mean time series were extracted from 300 regions of the seven-network Schaefer atlas. For each subject, a 300 × 300 CRBEN matrix was computed and decomposed using BrainSpace to obtain functional gradients, aligned via Procrustes analysis. Group differences were tested with two-sample t-tests controlling for age, sex, and education (Bonferroni-corrected, α = 0.05). Machine-learning classifiers were trained using gradient and demographic features, with robustness assessed by 1000 bootstrap resamples.

Results

Compared with NC, MCI showed reduced gradients in somatomotor, ventral-attention, and default-mode networks. AD showed further reductions versus MCI in somatomotor and default-mode networks, and versus NC in frontoparietal-control and default-mode networks (p < 0.05, corrected). Logistic regression achieved the highest accuracy (∼90%). Gradient flattening was prominent in temporal and occipital cortices, indicating reduced hierarchical organization.

Conclusions

CRBEN gradients demonstrate progressive loss of network complexity across the AD continuum and may provide sensitive biomarkers of functional disintegration.